Skip to main content
General

Vibe Writing: Borrowing the Developers' Word to Clarify the AI and Writing Debate

Developers have used the term vibe coding since February 2025. Brought into literature, it separates four ways of writing with or without AI and shows where the red line runs.

5 months ago9 min read

"There's a new kind of coding I call vibe coding, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists."
(Andrej Karpathy, on X, 2 February 2025)

Within a few months, the term vibe coding settled in among developers. Andrej Karpathy, one of the cofounders of OpenAI (the company that created ChatGPT), launched it one evening in February 2025 on X, half in jest, and it spread at a speed rare for a technical term. Merriam-Webster listed it as early as March among its "slang & trending" expressions, and the Collins dictionary made it its word of the year for 2025. Among developers, it now names a precise way of working: describe in plain language what you want, accept the code the model produces, run it and start again, without rereading the logic, without understanding it and sometimes without even opening the file.

The term hasn't reached literature yet, and it deserves to, because it names what the current debate between writers and AI can't put into words.

A debate going in circles

As soon as a novelist admits to using ChatGPT, two camps form. The purists believe no writer worthy of the name touches these tools, while the more open-minded see just one more tool, like the dictionary or a grammar checker. They aren't talking about the same thing, since the first confuse touching the tool with being replaced by it, and the second confuse occasional use with a writing engine.

The word they're missing is vibe writing.

Borrowing it from developers has a precise advantage: among them, vibe coding already carries a connotation. Nobody introduces themselves as a "vibe coder" in a meeting without a hint of irony, and the term describes the beginner who gives instructions to Cursor (an AI-assisted code editor) without rereading the result, ships the feature and has no idea what their own app does when a user clicks. The word thus carries an implicit judgment that separates the professional from the amateur without any need to explain it, and that's exactly what the literary debate lacks.

Four ways of writing

Once the word is in place, you have to say what it doesn't cover, and four levels emerge naturally.

1. Writing without AI. The writer types, crosses out and rereads alone. Many writers stand by this way of working, which remains entirely defensible: it's a discipline of the mind, and nobody should judge it as backward.

2. Augmented writing. AI enters the loop as a dictionary, a conversation partner or an outside view of the structure. The writer asks it for a synonym, a fact check or an opinion on a chapter's pacing, but writes every sentence by hand, and no generated text ever enters the manuscript. It's like the novelist who calls a doctor friend to check a pharmacology detail: the AI supplies information, and the prose stays the writer's own.

3. Co-writing. The writer keeps control of the direction, meaning the intent, the voice and the course of the story, but accepts that the AI proposes drafts, rephrasings or alternative versions of a scene, which the writer then rewrites in depth. Every sentence passes through their hands, so the final text is an openly mixed one in which the voice remains theirs. The method is demanding, and paradoxically it requires more judgment than writing without AI, because you constantly have to reject the average prose the model produces by default.

4. Vibe writing. The writer gives an instruction along the lines of "write me a chapter where Marie finds out her father lied to her, melancholy tone, 1,500 words," accepts the text as is or nearly so and moves on to the next prompt without rewriting anything. Their work comes down to choosing, accepting, rejecting or asking again, and the final manuscript is 80 or 90% generated prose that nobody has reworked. It's the exact equivalent of the vibe coder who ships an app without reading its code.

The red line: whose voice is it?

Today's debate traps itself by asking writers whether they use AI, yes or no. The question that really separates the practices is whose voice the final text carries.

The voice belongs to whoever rewrites. As long as the writer reworks each sentence, their signature survives, with its tics, its bursts, its deliberate lapses of taste and its rhythm, even when 60% of the first draft came from a model. Once they accept the raw text without running it through their own filter, their voice fades little by little, with nothing spectacular about it, until the prose becomes correct, readable and completely flat, which is the mark of a model far more than of an author.

Developers have known this line for two years. An experienced developer who works with Claude (Anthropic's assistant) all day remains an experienced developer, because they reject most suggestions, restructure the code and argue with the tool. A beginner doing vibe coding, by contrast, produces code that works without understanding it, and that breaks at the first unforeseen case. Writing works the same way, and there's no need for a new debate when the previous one has already been settled.

The purists' argument, and its weakness

Purists usually reply that writing is different, because code would be a utilitarian tool while a novel is an artistic act. The argument is appealing, but it holds up badly, for two reasons.

First, code also demands invention. Ask an experienced developer whether solving an architecture problem feels less creative to them than writing a line of crime-novel dialogue, and their answer will be clear. The two activities differ mainly in how visible their results are, since people read a novel, while they only notice a computer system when it breaks down.

Second, and above all, the argument doesn't survive the history of tools. The word processor didn't kill literature, grammar checkers didn't kill style and the proofreader didn't make the author disappear. Each generation greeted its tool with the same panic, and each time, demanding writers kept writing books worth reading, with one more tool in their kit.

AI differs from those tools because it can, in theory, replace the author instead of helping them. That's true, but only at the fourth level. At levels 2 and 3, it remains an instrument, like the tools that came before it, and mixing up the levels amounts to condemning the printed book because Gutenberg also made it possible to mass-print mediocre pamphlets.

An irony of the trade

Many publishers who publicly denounce AI-written manuscripts, and they're right to, are at the same time licensing their backlists to AI companies. HarperCollins, one of the Big Five American publishers, signed a deal with Microsoft in late 2024 worth about $5,000 per nonfiction title to train models. The author who signs the petition demanding "no AI books, ever" is thus asking a publisher that already licenses its books to Microsoft to hold a moral line that this publisher doesn't apply to itself.

You can sign the petition anyway, while looking at the debate with some distance, since each camp has its contradictions. The term vibe writing at least has the merit of clarifying what we're talking about, where the debate often settles for grand principles.

What this changes for a writer

Thinking about the debate in terms of these four levels changes three things for a writer questioning their own practice.

They can use AI without guilt, as long as they stay in charge of their text. A developer coding with Claude decides on every line, and a writer who asks ChatGPT to rephrase a paragraph before rewriting it occupies exactly the same position. The AI proposes, and now and then, somewhat at random and without intent, it produces a phrasing that unlocks a real idea, but it's the writer who decides what goes into the text. That's how a craftsman uses any tool.

There is a red line, and it has a name. If you ask a model to write 280 pages from a prompt and you sign the cover, you're vibe writing: you've bought a generation service far more than you've written a book. You can own that openly, since it's one way of producing books among others, as long as you say so instead of disguising it.

The tools sort themselves. Some online writing apps embrace vibe writing: Sudowrite, Novarrium and Squibler offer modes that generate an entire chapter. Others make the opposite choice of a discreet AI that can be switched off, only kicks in when asked and passes writers' texts to no outside model, like WriteControl, Extypis or Scribbook in French, and Dabble or Novlr in English. These tools differ less by being for or against AI than by the level of use they encourage.

A writer's job is therefore to choose the tool that matches the level they want to work at, far more than to choose between AI and nothing.

The word, to finish

Bringing vibe writing into literary vocabulary serves strategy as much as analysis. On the strategic side, the term arrives already loaded with the judgment the tech community put into it, so nobody needs to defend it. On the analytical side, it forces us to separate the levels and stop arguing in bulk.

A writer who uses a spell checker, or who has a paragraph rephrased before rewriting it by hand, remains a writer, whereas one who has an entire chapter generated and accepts it without reworking it is vibe writing.

Once that line is drawn, the debate about AI and writing can finally move forward, because we know what each person is talking about and at what level they stand, and we can judge a book on the work it contains rather than on its author's posture.


Sources

HU

Hubert Giorgi

Author

Write with Extypis

The writing studio novelists were missing

Outline, characters, places and objects, the links between them, mind map, scene board, export: everything lives in one place. Plus an AI that knows your story, whenever you want one.

Discover Extypis

Free to write and to export. No credit card asked.